Intelligent image-based in situ single-cell isolation

Quantifying heterogeneities within cell populations is important for many fields including cancer research and neurobiology; however, techniques to isolate individual cells are limited. Here, we describe a high-throughput, non-disruptive, and cost-effective isolation method that is capable of capturing individually targeted cells using widely available techniques. Using high-resolution microscopy, laser microcapture microscopy, image analysis, and machine learning, our technology enables scalable molecular genetic analysis of single cells, targetable by morphology or location within the sample.


Supplementary Note 1 -Laser-etched landmarks and registration evaluation
As a first step for every experiment, we etched 50x50 μm landmarks into poly-L-lysine coated slides (1 landmark per slide) using a microdissection microscope (Zeiss Axio Observer microscope with PALM MicroBeam manipulator). The landmarks fix positions that allow us to register image data between microscopes. They were designed to indicate the orientation to avoid any errors due to rotation of the coordinates. Unique barcodes may also be etched into the slide to identify samples.
The landmarks are first etched into the slides in a sample preparation step. In the high content microscope images, the landmarks are used to register the coordinates of the detected cells and contours selected for isolation.

Evaluation of registration error
Registration of the coordinate systems between the laser microdissection microscope and the high content microscope is imperfect. There are two main sources of error: 1) small scale differences between coordinate systems, 2) rotation, because it is impossible to consistently align the slide in different microscopes. We evaluate the accuracy of the landmark registration by etching two grid patterns, a fine pattern and coarse pattern, into a glass slide using the Zeiss PALM microscope with 63x objective. The grids were manually located in the second microscope and those locations were compared to the theoretic locations determined by using the landmarks.
The grid widths were 20 µm and 200 µm respectively. The aim was to simulate typical working conditions with the fine pattern (in which cells are densely crowded) and the coarse pattern aims to simulate larger areas of interest. The sample was then imaged using the PerkinElmer Operetta microscope, with 60x long working distance objective. A montage of the images is shown in Supplementary Figure 6.

Supplementary
We manually located all the corners of the grid patterns using the PerkinElmer Operetta microscope by zooming and panning. The measurements were made to pixel precision. A snapshot is shown in Supplementary Figure 7 The manually located grid corners were then compared to the theoretic locations determined by using one of the corners as a landmark. The registration errors are shown as More accurate registration can be achieved using more landmarks. With two landmarks, it is possible to accurately estimate translation and scale, and with three landmarks it is possible to accurately estimate translation, scale, and rotation. For our experiments, accuracy was deemed sufficient with a single landmark. In Supplementary Figure 9, we show the registration error fields for two and three landmarks. In conclusion, if the working area is near the landmark, a single landmark has sufficiently accurate registration. Two or three landmark registration significantly reduces registration error for applications where it is necessary. The nonlocal phase field energy used to model a gas of non-touching, non-overlapping near- * This model was originally published in 1 1 circular objects has the following form: where (un)primed functions are evaluated at (x ∈ D) x ∈ D ≡ D. The interaction function G : R 2 → R controls the range of interaction. The model can therefore be used in place of a classic active contour, but with the concomitant advantages of the phase field framework.
The above model is appropriate when object instances are well-separated, but it has a severe limitation: it cannot represent touching or overlapping object instances, because a phase field function represents subsets of D; and the nonlocal term in the energy, as well as generating the desired near-circular shapes, also causes object instances with small separation to have high energy.
To remove these difficulties, a multi-layer version of the above model was developed in 3 .
where is the number of layers. The energyẼ f (φ) of the model is simply the sum of energies of independent layers extended with a term that penalizes overlapping pairs of object instances by an amount proportional to overlap area: where κ controls the overlap penalty. This model solves the two issues mentioned above. Overlapping object instances can now be represented by appearing in different layers, and the inter-object repulsion is now removed because it is energetically favourable for nearby objects to be represented 2 on different layers, thus incurring no energy penalty.
Datam model for overlapping cells. We model the image intensities in the background and in the single-object foreground (i.e. a cell) as having fixed (but different) means and variances, leading to Gaussian distributions with independent pixel intensities by maximum entropy. When several objects overlap, we model the resulting image as the sum of the intensities from the background and each of the overlapping objects, so that the resulting model is again Gaussian with independent pixels, but with mean and variance equal to the sum of the means and variances of the background and the objects.
, which represents the number of overlapping objects at each point ( 1 ). Then the likelihood energy is where I is image intensity; µ − and σ 2 − are the mean and variance of the background; and ∆µ and ∆σ 2 are the changes in mean and variance brought about by each new overlapping object.
Pre-processing Iit was shown 1 that the 'multi-layer gas of circles' model resutls more accurate segmentation by using initial seeds inside the object to be segmented. To find nuclei centers in "difficult images" an A-Trous wavelet transform based spot detection were used 4 . The computational complexity is proportional to the number of 'gas of circles' layers, we have to minimize the number of these layers. For that we chose a graph/map coloring method as follows. Let each seed centroid be a node in a graph. An edge connects two nodes if they are close (this distance was set 3 to 3.5 times the preferred object radius). After building the graph we can color the nodes with 4 color to provide the none of connected node pairs has the same node color.
Post-processing To avoid 'degenerate' configurations the following was applied: • Discard embedded smaller objects. All objects that is completely surrounded by any other object on an other layer, is removed.
• Merge fully overlapping objects. If the Jaccard index of two objects from different layers is bigger than 0.80, they are merged.